File size: 5,314 Bytes
655ac3b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
---
license: other
license_name: composite
pretty_name: VisDocAgentBench
language:
- en
tags:
- visual-document-retrieval
- agentic-search
- multimodal-retrieval
- scientific-documents
size_categories:
- 1K<n<10K
configs:
- config_name: queries
  data_files:
  - split: test
    path: benchmark/queries.jsonl
- config_name: evaluator_annotations
  data_files:
  - split: test
    path: benchmark/evaluator_annotations.jsonl
- config_name: documents
  data_files:
  - split: corpus
    path: corpus/documents.jsonl
- config_name: pages
  data_files:
  - split: corpus
    path: corpus/pages.jsonl
---

# VisDocAgentBench

VisDocAgentBench is a closed-corpus benchmark for visually rich document retrieval. It contains 120 natural-language queries over 2,375 rendered pages from 100 scientific documents. Queries are evenly divided among direct, one-bridge, and two-bridge evidence structures, with one answer page per query.

[[Paper](https://arxiv.org/pdf/2608.17889)] [[Code](https://github.com/hulx2002/VisDocAgentBench)] [[Project page](https://hulx2002.github.io/VisDocAgentBench)]

## Contents

```text
benchmark/
├── queries.jsonl
├── evaluator_annotations.jsonl
└── topics.json
corpus/
├── documents.jsonl
├── pages.jsonl
└── pages/<document_id>/<page_id>.png
dataset_info.json
LICENSES.md
```

- `queries.jsonl` contains the query text and topic identifier presented to a retrieval system.
- `evaluator_annotations.jsonl` contains the answer page, evidence level, and ordered latent support pages used by the evaluator and controlled analyses.
- `documents.jsonl` records the exact arXiv version, bibliographic metadata, source URLs, source license, page count, and image-availability status for every document.
- `pages.jsonl` defines all 2,375 page identifiers, document membership, one-based page indices, rendering dimensions, and expected local paths.
- `topics.json` describes the ten corpus topics.

The standard agent harness reads the query text but does not expose answer or support annotations to the planner.

## Dataset Statistics

| Statistic | Count |
|---|---:|
| Documents | 100 |
| Rendered pages | 2,375 |
| Queries | 120 |
| Direct queries | 40 |
| One-bridge queries | 40 |
| Two-bridge queries | 40 |
| Unique answer pages | 120 |
| Directly included page images | 1,469 |
| Locally reconstructed page images | 906 |

## Loading the Metadata

Each JSONL component has a distinct schema and can be loaded independently:

```python
from datasets import load_dataset

queries = load_dataset(
    "hulx2002/VisDocAgentBench",
    "queries",
    split="test",
)
annotations = load_dataset(
    "hulx2002/VisDocAgentBench",
    "evaluator_annotations",
    split="test",
)
pages = load_dataset(
    "hulx2002/VisDocAgentBench",
    "pages",
    split="corpus",
)
```

The complete snapshot can be downloaded with the code repository:

```bash
python scripts/download_data.py
```

## Reconstructing the Complete Corpus

The repository includes 1,469 rendered page images whose source licenses permit redistribution. The remaining 906 rows remain in `corpus/pages.jsonl` with `image_included=false`. The exact source version and PDF URL are recorded in `corpus/documents.jsonl`.

After cloning the [code repository](https://github.com/hulx2002/VisDocAgentBench) and downloading this dataset into `data/`, reconstruct the omitted pages with:

```bash
python dataset_tools/download_and_render.py
python dataset_tools/validate_dataset.py --require-complete-corpus
```

The script downloads each specified arXiv version and renders it at 144 DPI. It verifies the expected page count and dimensions before accepting the reconstructed corpus.

## Evaluation

Systems return a ranked list of page identifiers or the opaque page handles assigned by the released agent harness. The evaluator reports Recall@1/3/5/10 and MRR@10 overall and by evidence level. Missing or invalid rankings stay in the 120-query denominator and score zero.

Evaluation code, baseline implementations, deterministic preprocessing, and the full agent tool interfaces are available in the [code repository](https://github.com/hulx2002/VisDocAgentBench). Generated predictions, traces, OCR caches, embeddings, and model weights are not included in this dataset repository.

## Licensing

Benchmark-authored queries, evaluator annotations, topics, and corpus metadata are licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Source-document page images retain the license of their source document. The license URL and attribution metadata for every source are recorded in `corpus/documents.jsonl`.

Page images are included only for 55 CC BY 4.0 documents and 4 CC BY-NC-SA 4.0 documents. Sources under the arXiv nonexclusive distribution license or CC BY-NC-ND 4.0 are represented by metadata and local reconstruction instructions, not redistributed images. See [LICENSES.md](LICENSES.md) before reusing source pages.

## Citation

```bibtex
@article{hu2026visdocagentbench,
  title={VisDocAgentBench: Benchmarking Agents for Visually Rich Document Retrieval},
  author={Hu, Lexiang and Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Li, Yikang and Zhang, Fuwei and Wang, Yisen and Lin, Zhouchen},
  journal={arXiv preprint arXiv:2608.17889},
  year={2026}
}
```